A0864
Title: Functional dynamic tensor decomposition with application to modelling yield curves across multiple countries
Authors: Xinyu Wang - The University of Hong Kong (Hong Kong) [presenting]
Qin Fang - the University of Sydney (Australia)
Xinghao Qiao - The University of Hong Kong (Hong Kong)
Abstract: A novel tensor-based framework is proposed for modeling multivariate functional time series, formulated as an order-3 tensor with feature, functional, and temporal modes. The framework incorporates two functional tensor decomposition methods: (i) a Tucker decomposition model, where the dynamics are driven by a low-dimensional core tensor (viewed as a matrix-valued time series), and (ii) a CP decomposition model, which approximates the series as a sum of rank-1 tensors. Both approaches preserve the inherent matrix or tensor structure, enabling effective dimension reduction along the feature and functional modes simultaneously. A model selection procedure is further developed to identify the more suitable decomposition for a given dataset. Simulation studies demonstrate that the proposed methods offer improved performance in capturing complex dependency structures and enhancing interpretability.
